Default factor levels when coercing from logical

[alejandroschuler]

Hi folks, I'm not sure if forcats is the right place to raise this issue, so apologies in advance if this should go elsewhere.

I find it curious that factor(c(TRUE, TRUE)) produces a factor with a single level: TRUE. If we are to imagine that factors are categories where we know all of the levels a-priori, wouldn't it make sense if logicals converted to factors preserved that a-priori knowledge of the two levels TRUE and FALSE (with a consistent order) regardless of whether or not one of those two levels does not appear in the data? It's a trivial fix using levels=c("TRUE", "FALSE"), but I find it to be so philosophically obvious that this should always happen that it should warrant default behavior. Perhaps something along the lines of:

fct.default = base::factor
fct.logical = function(lgl, ...) base::factor(lgl, levels=c("TRUE", "FALSE"))
fct = function(x, ...) UseMethod("fct")

This is a sometimes a minor frustration for me when using tidymodels packages (and caret before that) which require a factor outcome for classification problems. Every now and again I end up in a corner case where I get errors like:

 Error: In metric: `yardstick::sens`
`truth` and `estimate` levels must be equivalent.
`truth`: TRUE, FALSE
`estimate`: TRUE

[hadley]

Makes sense, but what about NA? There are three possible values of a logical vector after all.

[alejandroschuler]

NA behavior is the same as with base::factor.

> fct(c(TRUE, FALSE, NA))
[1] TRUE  FALSE <NA> 
Levels: TRUE FALSE
> base::factor(c(TRUE, FALSE, NA))
[1] TRUE  FALSE <NA> 
Levels: FALSE TRUE

As is the case with other factors, NAs are represented as NAs, not as an explicit level. I would say this is the desired default behavior because although NA is a possible value that the variable can take in the data, it is not a value the variable can take epistemologically. If that behavior is desired, fct_explicit_na does the job as usual.

[alejandroschuler]

Perhaps the best place to implement this in forcats would be as a method of as_factor:

as_factor.logical = function(lgl, ...) base::factor(lgl, levels=c("TRUE", "FALSE"), ...)

Although this would not be consistent with leveling in order of appearance.

[alejandroschuler]

However, I don't think the justification for sorting by order of appearance (locale variation) applies to logicals (or numerics). So maybe

as_factor.logical = function(lgl, ...) base::factor(lgl, levels=c("TRUE", "FALSE"), ...)
as_factor.default = function(x, ...) as.factor(x, ...)

What I'm not sure about is whether first level of the logical factor should be FALSE to be consistent with the first level of something like as_factor(c(1,0)) being 0 (implying yardstick.event_first=FALSE should be default). It may be better to have TRUE come first to explicitly break the connection between logicals and integers (implying yardstick.event_first=TRUE should be default, as it is).

[hadley]

I think the default levels would match the default ordering, so they'd be c("FALSE", "TRUE")